Maternal education, parental investment and non-cognitive skills in
rural China
Jessica Leight∗1 and Elaine M. Liu2
1
2
Williams College Department of Economics
University of Houston Department of Economics and NBER
April 25, 2016
Abstract
The importance of non-cognitive skills in determining long-term human capital and labor
market outcomes is widely acknowledged, but relatively little is known about how educational
investments by parents may respond to non-cognitive skills early in life. This paper evaluates
the parental response to variation in non-cognitive skills among their children in rural Gansu
province, China, employing a household fixed effects specification; non-cognitive skills are defined as the inverse of both externalizing challenges (behavioral problems and aggression) and
internalizing challenges (anxiety and withdrawal). The results suggest that on average, parents
invest no more in terms of educational expenditure in children who have better non-cognitive
skills relative to their siblings. However, there is significant heterogeneity with respect to maternal education; less educated mothers appear to reinforce differences in non-cognitive skills
between their children, while more educated mothers compensate for these differences. Most
importantly, there is evidence that these compensatory investments lead to catch-up in noncognitive skills over time for children of more educated mothers.
1
Introduction
In recent years, both research and policy debates have placed increasing emphasis on the importance
of non-cognitive skills in determining long-term economic outcomes. Data primarily from industrialized countries has suggested that non-cognitive skills have a large impact on adult economic
∗
Thanks to Doug Almond, Janet Currie, Flavio Cunha, Rebecca Dizon-Ross, Benjamin Feigenberg, Paul Frijters,
Paul Glewwe, Lucie Schmidt, Kevin Schnepel, Chih-Ming Tan, and Andy Zuppann for helpful conversation and suggestions, and to seminar and conference participants at the University of Minnesota - Applied Economics, Williams,
the ASSA meetings, the Chinese Economists’ Society conference, the Western Economic Association conference, the
University of North Dakota and the University of Montreal - CIREQ conference for feedback. Previous drafts were
circulated under the title “Maternal bargaining power, parental compensation and non-cognitive skills in rural China.”
Keywords: non-cognitive skills, parental investment, intrahousehold allocation. JEL codes: I24, O15, D13
1
welfare, measured as earnings and labor productivity (Heckman and Rubinstein, 2001; Heckman,
Stixrud and Urzua, 2006; Cunha, Heckman, Lochner and Masterov, 2006; Carneiro, Crawford and
Goodman, 2007). There are many causal pathways through which stronger non-cognitive skills may
lead to improved educational and economic outcomes: individuals with enhanced skills are more
likely to be persistent in achieving strong academic outcomes or building professional expertise,
may be more resilient in the face of setbacks, or may be better able to forge useful professional
relationships. One particularly important channel, however, is the relationship forged much earlier
in life between children and their parents.
Variation in non-cognitive skills may affect parental investments in several ways. First, this
variation could alter the weight that a parent places on a child’s welfare. Parents could favor a
child with stronger non-cognitive skills with whom they forge a stronger relationship, or a child
with weaker non-cognitive skills if he seems to require more nurturing. Second, even if the weight
parents place on their children’s welfare is unchanged, non-cognitive skills will affect children’s
future income, and may affect the returns to human capital investment in a given child. Depending
on whether parents emphasize efficiency or equality, they may then invest more or less in human
capital development for children of varying levels of non-cognitive skills.
Ultimately, we may observe compensatory patterns of investment, in which parents invest more
in a relatively weaker child, or reinforcing behavior, in which parents invest more in a stronger
child; the question of which pattern is dominant is an important question in family economics
going back to Becker. Moreover, given that a number of evaluations have found that targeted early
intervention can affect children’s non-cognitive skills,1 understanding whether these interventions
crowd parental investment in or out may be a useful contribution to the ongoing policy debate.
The objective of this paper is to analyze whether non-cognitive skills measured in childhood and
adolescence have a significant impact on the within-household allocation of educational expenditure
among households in rural Gansu province, China. We employ a panel dataset that provides a
detailed set of outcome measures for a large cohort of children in one of the poorest provinces in
China; non-cognitive skills are measured via direct surveys of the sampled children, and defined
as the inverse of externalizing challenges (behavioral problems and aggression) and internalizing
challenges (withdrawal and anxiety). Focusing on a sample of two-children families, our primary
specification examines how parents respond to differences in non-cognitive skills conditional on
household fixed effects, and whether this response varies based on the characteristics of the parents.
Our results suggest that while parents are not responsive to differences in non-cognitive skills
on average—neither reinforcing nor compensating for these differences—there is significant heterogeneity with respect to characteristics of the parents, and particularly the mother. Households
with more educated mothers show evidence of significantly more compensatory investment compared to households with less educated mothers. For a child who exhibits non-cognitive skills one
1
The results from the Perry preschool study as reported in Schweinhart et al. (2005) are among the best known
in this respect.
2
standard deviation lower than his sibling, an increase in maternal education from the 25th to the
75th percentile, or from one to six years of education, would result in an increase in discretionary
educational expenditure (comprising all educational expenditure excluding tuition) directed to this
child of nearly 35%; there is no comparable effect observed for tuition. In addition, there is very
little evidence of comparable heterogeneity with respect to the education of the father.
One potential challenge faced in this analysis is that non-cognitive skills may in fact be partly
an outcome of previous parental investment, rather than an endowment; if there is some serial
correlation in parental investment, this will generate bias toward the detection of a reinforcing
pattern of expenditure. We address this challenge in several ways. First, we demonstrate that
bias on the interaction effect including maternal education will only arise specific assumptions that
do not seem to be supported by the empirical evidence. Second, we present evidence that the
results are robust to several alternate specifications, including the use of earlier measurements of
non-cognitive skills and directly controlling for past expenditure.
This observed pattern of heterogeneous response—more compensation in households with a
more educated mother—could be consistent with a number of possible channels. More educated
mothers may simply be better able to recognize the non-cognitive deficits in their children and
to compensate appropriately. Maternal education may be a proxy for income or other household
characteristics, and higher-income households may have a preference for intrahousehold compensation. Alternatively, women (or more educated women in particular) may have a preference for
compensatory investment, and the ability to impose these preferences within the household.
Further exploration suggests that the first two channels are not particularly salient in this
context: there is little evidence consistent with parental learning or higher income leading to a
compensatory response by more educated mothers. However, it does seem that more educated
mothers may have both a greater preference for compensatory investment and higher bargaining
power, enabling them to exert more influence over the allocation of expenditure within a household.
In the final section of the paper, we analyze whether this variation in compensatory vis-a-vis
reinforcing behavior results in catch-up in non-cognitive skills over time for children in households
with more educated mothers—where struggling children receive greater investment—compared to
children in households with less educated mothers. Analyzing longitudinal data observed for the
first-born child over time, we find evidence of significantly greater catch-up in non-cognitive skills
between ages 9-12 and ages 17-20 for children of more educated mothers. In other words, the correlation between maternal education and non-cognitive skills becomes more pronounced as children
age: in the first wave, the observed cross-household correlation between a dummy for a mother of
high maternal education and the percentile measures of non-cognitive skills is essentially zero. In the
second wave, the correlation has increased in magnitude to .028, and by the third wave, .151. This
result highlights the importance of maternal education on the formation of non-cognitive abilities,
and contributes to the literature on the importance of maternal education on child development
3
(Carneiro et al., 2013; Currie, 2009). If the economic returns to non-cognitive skills are significant,
this is a channel through which inequality across households can widen over time.
Our paper contributes to several related literatures on intrahousehold allocation and human capital investment. First, there is an extensive literature examining parental responses to differences
in children’s endowment; Almond and Mazumder (2013) provide a recent review. The evidence has
been mixed. Bharadwaj et al. (2013b), Royer (2009), and Almond and Currie (2011) find little or
no evidence of either compensatory or reinforcing behavior. Akresh et al. (2012), Rosenzweig and
Zhang (2009), Frijters et al. (2013), Almond et al. (2009), Adhvaryu and Nyshadham (2014) and
Aizer and Cunha (2012) find parents exhibit reinforcing behavior in Burkina Faso, China, Sweden,
Tanzania, and the United States, while Del Bono et al. (2012) find evidence of compensatory behavior in breast-feeding decisions and birth weight and Black et al. (2010) provide some indirect
evidence of compensatory behavior in a robustness check. Bharadwaj et al. (2013a) find compensatory investment with respect to initial health comparing across siblings, but comparing across
twins, there is no evidence of either compensatory or reinforcing behaviors. Leight (2014) finds
evidence of compensatory behavior with respect to height-for-age using the same sample as this
paper. This literature, however, focuses primarily on parental responses to children’s health endowment and cognitive ability. Our paper is one of the first papers that examines whether parents
respond to children’s non-cognitive skills in the allocation of human capital investment.2
Second, our paper finds that maternal education plays an important role in determining the allocation of parental investment. This is similar to the results reported in Hsin (2012) and Restropo
(2016), who conclude that households with less educated mothers generally exhibit reinforcing investment behavior, while households with more educated mothers exhibit compensatory behavior.
However, neither of these papers examine the father’s education level; instead, they analyze maternal education as a proxy for household socioeconomic status. In light of these findings, a recent
review by Almond and Currie (2011) suggests that the observed pattern could be due to credit
constraints in low socioeconomic status households, or a high elasticity of substitution between
consumption and human capital investment in these households.
In our context, we are able to rule out these two channels given the absence of any evidence
of heterogeneity with respect to the father’s education and household income. Rather, we present
additional results that suggest differential preferences and bargaining power for more educated
women are the primary channels for the differential compensatory response in households with
more educated mothers.
Third, there is growing evidence that early intervention and investment can mitigate initial
deficits in children’s endowments (Bhalotra and Venkataramani, 2016; Adhvaryu et al., 2015; Bleakley, 2010; Cunha and Heckman, 2007; Cunha et al., 2010; Gould et al., 2011; Almond et al., forth2
The only other relevant paper is Gelber and Isen (2013). While it is not the focus of their work, they report in
their appendix that parents respond positively to a child with greater observed non-cognitive abilities, but they do
not further investment the heterogeneity by mother’s education.
4
coming; Kling et al., 2007). However, this literature primarily focuses on evidence of mitigation in
children’s health and cognitive outcomes. To the best of our knowledge, we are the first to show
evidence of catch-up in non-cognitive skills for children as a result of parental investment, especially
among more educated mothers.
The remainder of the paper proceeds as follows. Section 2 describes the data. Section 3
describes the empirical strategy and the primary results, and Section 4 presents robustness checks
and evidence about the relevant channels. Section 5 examines the longitudinal evidence about
persistence of non-cognitive skills over time, and Section 6 concludes.
2
Data
The data set used in this paper is the Gansu Survey of Children and Families (GSCF), a panel
study of rural children conducted in Gansu province, China. Gansu, located in northwest China,
is one of the poorest and most rural provinces in China. The description of the data here draws
substantially on the description in Leight (2014).
The first wave of the GSCF was conducted in 2000, and surveyed a representative sample of
2,000 children aged 9–12 in 20 rural counties, supplementing these surveys with additional surveys
of mothers, household heads, teachers, principals, and village leaders. These children are denoted
the “index children.” All but one of the index children have complete information in the first wave.
The second wave, implemented in 2004, re-surveyed the first sample of children at age 13–16
and also added a survey of their fathers. 1,872 children, or 93.6% of the original sample, were
re-interviewed in the second wave. In addition, surveys were added of the eldest younger sibling of
the index child. These additional children are denoted “younger siblings.” Surveys are conducted
directly with the younger siblings, as well as with their homeroom teachers; in addition, mothers
and fathers report limited supplementary information about the younger siblings.
In early 2009, a third wave of surveying was conducted, re-interviewing the index children
during Spring Festival, a period at which many of them had returned to their natal villages. In
cases where the sampled individual was not available, parents were asked to provide information
about their child’s education and employment status. 1,437 individuals, or 72% of the original
sample, were interviewed directly in this wave, and information was collected in parental interviews
for an additional 426 sample children.
The household surveys in waves one (2000) and two (2004) included extensive questions about
schooling outcomes, household expenditure on education for each child, child time use, time investments in education by parents and teachers, and child and parental attitudes, as well as more
standard socioeconomic variables. The index children also completed a number of achievement and
cognitive tests. Younger siblings also completed these tests in wave two.
In addition, each wave of data collection included survey questions posed to the sample children that were designed to measure their non-cognitive skills. In the first and second waves, the
5
survey measured both internalizing and externalizing behavioral challenges: the former refers to
intra-personal problems (e.g., withdrawal and anxiety), and the latter to inter-personal problems
(destructive behavior, aggression, and hyper-activity). Both measures of non-cognitive skills are
constructed by recording the respondent’s agreement or disagreement with a series of statements
and then applying item response theory (IRT) to generate internalizing and externalizing scores.
The measures are identical across waves one and two, and the scores are standardized to have a
mean of zero and a standard deviation of one. In the third wave, a Rosenberg self-esteem index
and a depressive index were measured. Further detail about the construction of the non-cognitive
skills measures can be found in Glewwe, Huang and Park (2013).
Non-cognitive skills of the younger siblings were measured only in wave two. For ease of interpretation, the primary non-cognitive skills measures employed here (the externalizing and internalizing
indices) have been inverted; in the original index, a higher value indicates more challenges, but in
our index a higher value indicates better non-cognitive skills.
In this paper, we will focus on a subsample of the families in the survey: those with two children
in the household where both children have reported measurements for non-cognitive and cognitive
skills in the second-wave survey. If the index children and the younger sibling are the only children
in the household, then the surveys provide a complete overview of parental allocations and child
endowment. Complete data is available for 388 families drawn from 90 localities in 20 counties,
and these households constitute the relevant subsample. In our sample, only 6.5% of households
have one child. The remaining households are excluded because the index child has two or more
siblings, or has one older sibling for whom non-cognitive ability is not reported. In the robustness
checks, we will also present results employing a slightly larger sample including households where
these two children (the index child and the younger sibling) are part of a larger family.3 Figure 1
summarizes the structure of the sample, including the years in which data is collected, the children
that are observed in each wave, and their age at the point of data collection.
Panel A of Table 1 reports summary statistics for the subsample of two-children families and
the overall sample for key demographic indicators of interest, as well as a t-test for equality between
the two means; the covariates reported are measured in the second wave of the survey, the wave
in primary use here. It is evident that there are no significant differences in income or parental
education between the sample and the subsample. However, households in the subsample are
slightly younger and have younger children. This primarily reflects the exclusion of larger families
or families in which the index child is the younger child; these families are generally headed by older
parents. Importantly, there are also no significant differences between the reported non-cognitive
and cognitive skills in the second wave for the index children in the full sample and the subsample.
3
While China’s One-Child Policy was in effect during the period in which these children were born, many rural
households could nonetheless have two children legally under various exemptions to the policy (Gu et al., 2007). It is
not possible using this dataset to accurately identify for each household whether it was in technical compliance with
the policy.
6
The dependent variable of interest is educational expenditure per child per semester, reported
by the head of household in six categories: tuition, educational supplies, food consumed in school,
transportation and housing, tutoring, and other fees.4 Each household separately reports expenditure for each child in each of these categories. Discretionary expenditure is defined as the sum
of all expenditures excluding tuition. Summary statistics for average expenditure per child for the
subsample of families analyzed can be found in Panel B of Table 1. Total educational expenditure averages around 360 yuan per child per semester; an average of 20% of household income is
allocated to educational expenditure in total for both children.
We focus on educational expenditure given that it is the primary form of child-specific expenditure reported in this dataset. The only other type of child-specific expenditure reported is medical
expenditure over the past year; only 25% of households report any positive medical expenditure
for either child over the past year, and unsurprisingly this expenditure is highly correlated with
reported illness (i.e., it is reasonable to assume that very little corresponds to preventive care).
Given that we are primarily interested in human capital investments with long-term returns, we do
not focus on medical expenditure.
3
Empirical strategy and results
3.1
Empirical strategy
Our empirical strategy entails evaluating whether parental expenditure on education for children
is correlated with measures of non-cognitive skills, conditional on household fixed effects.5 In other
words, our primary specification identifies whether parents are more likely to invest in a child who
has stronger non-cognitive skills relative to a sibling.
The child’s observed non-cognitive skills will be denoted N cogihct , for child i in household
h, living in county c and born in year t: the non-cognitive variables employed will include the
externalizing and internalizing index, as well as a summary measure of non-cognitive skills that is
the mean of the two indices.
All non-cognitive indices have been standardized to have means equal to zero and standard
deviations equal to one, and in order to facilitate interpretation, the indices have been inverted
such that a higher value is indicative of stronger non-cognitive skills. The dependent variable,
educational expenditure, is denoted Yihct . The specification includes household fixed effects ηh ,
year-of-birth fixed effects νt , and a vector of child-level covariates Xihct , yielding the following
equation. Child-level covariates include gender, the sibling’s gender, birth parity (i.e., whether a
4
In China, textbook fees are mandatory and levied as part of the overall tuition, and here they are likewise reported
in the tuition category. Educational supplies is supplies other than textbooks.
5
Given that our primary interest is examining how parents allocate resources between children with different
non-cognitive skills, it is necessary to include family fixed effects since we are examining within household variation.
Also, the use of fixed effects addresses the challenge of unobserved household-level heterogeneity.
7
child is first-born or second-born), height-for-age, reported grades in school, and scores on gradespecific achievement tests administered in math and Chinese. The inclusion of birth parity is
particularly important, given the evidence presented by Black et al. (2005a) suggests that birth
order is an important determinant of children’s outcomes.
This specification is estimated with and without interactions with parental education Shct .
Standard errors will be clustered at the county level in all specifications.6
Yihct = β1 N cogihct + β2 N cogihct × Shct + Xihct + νt + ηh + ihct
(1)
The identification assumption for this family of specifications requires that non-cognitive skills are
uncorrelated with other unobservable variables that determine parental allocations. This assumption would be violated, for example, if parents invest more in a favored child, who is then observed
to have stronger non-cognitive skills.
In order to present some preliminary evidence about the relationship between non-cognitive
skills and child characteristics conditional on household fixed effects, the following specifications
can be estimated, regressing the internalizing and externalizing indices on child covariates Xihct ,
conditional on household fixed effects.
N cogihct = β1 Xihct + ηh + ihct
(2)
The results can be found in Table 2: Panel A reports the correlations with sibling parity, age,
gender, and grade level, and Panel B reports the correlations with various measures of the child’s
endowment. Interestingly, in Panel A there is no evidence of any significant correlation between the
internalizing index and any child characteristic. However, for the externalizing index we observe
that non-cognitive skills are weaker for second-born children, younger children, boys and children
enrolled in lower grades in school. There is, of course, a high degree of correlation among these
covariates: second-born children are on average younger, enrolled in lower grades, and more likely
to be boys.7 Columns (9) and (10) show the results of a multiple regression including all four
covariates; there is some evidence here that the most robust correlations are between gender and
grade level and the externalizing index.
Panel B shows that there is little evidence of significant correlations between non-cognitive skills
and other measures of endowment: specifically, height-for-age, and various measures of cognitive
skills. This includes the child’s grades reported in the last academic year in math and Chinese,
and their score on a grade-specific achievement test. The only exception is a significant correlation
between the internalizing index and the achievement test score.
In light of these results, the primary specifications all include year-of-birth fixed effects as well as
6
We regard clustering at the county level as a conservative strategy for inference. Our subsample includes 20
counties and 90 villages; our results are also consistent if we cluster at the village level.
7
The implications of gender selection for this analysis will be explored in greater detail in Section 4.1.
8
controls for gender, sibling’s gender, sibling parity, height-for-age, and all three reported measures
of cognitive skills as measured in the second wave, contemporaneously with non-cognitive skills.
The inclusion of grade fixed effects is more complex; given that the primary dependent variable of
interest is educational expenditure, grade level can plausibly be considered an outcome. However,
the primary results will also be robust to the inclusion of grade fixed effects.
Given that the primary specifications are estimated conditional on household and year-of-birth
fixed effects, it is also useful to examine how much variation in the child characteristics of interest
is observed within a given household and within a given birth year. Table A1 in the Appendix
reports the R-squared in a simple regression including only the specified fixed effects as explanatory
variables and the specified measure of non-cognitive or cognitive skills as the dependent variable.
In general, between 50% and 60% of the variation in non-cognitive skills is explained by household
fixed effects, suggesting there is still considerable within-household variation.
3.2
Primary results
Table 3 shows the results of estimating equation (1) without the interaction terms with parental
education. The objective is to test whether parental allocations of educational expenditure are
responsive, on average, to variation in non-cognitive skills between siblings; the measures of expenditure include total expenditure, discretionary expenditure (the sum of all expenditure excluding
tuition), and six individual categories. Enrollment is universal among the subsample of interest,
and thus school enrollment is not reported as an outcome. The results show coefficients that are
small in magnitude, varying in sign, and generally insignificant. This suggests that parents are
neither systematically compensating children with weaker non-cognitive skills, nor systematically
reinforcing these differences. A similar pattern is observed if we estimate this regression in the
cross-section without household fixed effects.
In light of this pattern, we then examine whether parents who themselves have certain characteristics are more likely to respond to measured differences in the non-cognitive skills of their children.
The most obvious relevant characteristic is education, particularly given that recent work from
Hsin (2012) and Restropo (2016) suggests that maternal education is an important determinant
of parental allocation of expenditure. While the average level of education reported is relatively
low—four years for mothers and seven years for fathers—there is considerable variation. Around
75% of mothers report completing at least one year of formal schooling, and 17% report completing
junior high school. For fathers, around 90% report completing at least one year of formal schooling,
and 10% report completing senior high school.
Figure 2 shows histograms of the distribution of both maternal and paternal education. The
correlation between maternal and paternal education is positive, but low in magnitude (around .3).
Unsurprisingly, the education of both parents is also positively correlated with income and other
measures of household wealth.
9
In addition, Figures 3a and 3b show the distribution of intrahousehold (between-sibling) differences in non-cognitive skills and in normalized expenditure residuals for households at different
levels of maternal education. We can observe that the mean absolute difference in non-cognitive
skills between siblings is around .75 standard deviations, and this is roughly constant across households of different levels of maternal education. Normalized expenditure residuals are calculated by
regressing expenditure on child characteristics (gender, birth parity, cognitive skills, and height-forage), generating the residuals and standardizing them to have mean zero and standard deviation
one. The mean absolute difference is around .4 standard deviations, and this seems to be larger at
higher levels of maternal education.
To identify whether parents who are more educated respond differentially to differences in noncognitive skills, we then re-estimate equation (1) including interaction terms between non-cognitive
skills and parental education. Again, the specification includes controls for a wide range of child
characteristics including gender, sibling parity, and cognitive skills, and household and year-of-birth
fixed effects. We also include the interactions of gender, age, sibling parity, height-for-age and the
achievement test score with the specified measure of parental education. The results analyzing
variation in parental response to children’s non-cognitive skills with respect to parental education
are reported in Table 4.
We observe a robust pattern in which households in which mothers have low levels of education
(roughly speaking, fewer than three years of schooling) provide more expenditure to children with
better non-cognitive skills, reinforcing the pre-existing differences, while households with more
educated mothers seem to engage in compensatory behavior, providing more expenditure to children
with worse skills. This is evident in the negative coefficients on the interaction term between noncognitive indices and maternal education.8 The coefficients on the interaction terms for other child
characteristics are not reported in Table 4 for concision, but are reported in the on-line appendix
for a simple specification using a summary measure of non-cognitive skills.9 For cognitive skills
as captured by a summary achievement test score, β1 is positive though insignificant, suggestive
of a reinforcing allocation of investment, and there is no evidence of heterogeneity with respect to
maternal education; more discussion about this pattern is provided in Section 4.2.
In general, there is no compensatory effect observed for tuition, consistent with the intuition
that tuition is not easily manipulable by parents; though the coefficients are negative, they are
statistically insignificant and small in magnitude.10 For example, comparing the estimated coefficients for the interaction effect in Columns (2) and (3) in Panel A, we can observe that relative
to the mean of the dependent variable, the magnitude of the coefficient on tuition in Column (3)
8
Aizer and Cunha (2012) find that the degree of parental reinforcing behaviors increases with family size. However,
our finding cannot be explained by variation in family size since the sample is restricted to only two-child households.
9
The on-line appendix can be found on both authors’ homepages.
10
Approximately 12% of students attend schools that are are private or publicly assisted private institutions;
accordingly, it is possible that there is some variation in tuition, and some potential for parents to select a highertuition school for their children.
10
is around 10% of the magnitude of the coefficient on discretionary expenditure in Column (2). A
similar pattern is observed in Panel C.
The interaction terms for paternal education, on the other hand, are smaller in magnitude,
heterogeneous in sign, and generally not statistically significant. Panels B and D report p-values
testing the equality of the estimated coefficients on the interaction terms for maternal and paternal
education for the internalizing and externalizing indices, respectively; this test is implemented by
estimating the two specifications simultaneously in a seemingly unrelated regression framework.
These coefficients are denoted β2m and β2f , where β2m refers to the coefficient on the interaction
term for maternal education, and β2f refers to the analogous coefficient for paternal education.
While the imprecision in the coefficients on the paternal education interaction term does not allow
us to reject equality in all cases, we can reject the hypothesis that the coefficient on the maternal
and the parental education interaction terms is equal in both specifications employing total and
discretionary expenditure as the dependent variable, as well as in several additional specifications.
We also report results in both panels from a joint test of the hypotheses β1m = β1f and β2m = β2f ,
and find generally parallel results.
The magnitudes of the implied effects are also substantial. For example, consider a child who
exhibits non-cognitive skills as measured by the internalizing index that are one standard deviation
lower than his sibling. The coefficient on the interaction term suggests that an increase in maternal
education from the 25th to the 75th percentile, or from one to six years, would result in an increase in
discretionary educational expenditure for this child compared to his sibling of 35%. The magnitudes
are similar for the coefficients estimated for the externalizing index in Panel C.
4
Robustness checks and channels
4.1
Robustness checks
Alternate specifications First, we verify the primary results are robust to a number of additional specifications. For concision, in the robustness checks we employ an average non-cognitive
index that is the mean of the internalizing and externalizing indices. The observed pattern is
consistent if grade fixed effects are included, if the non-cognitive variables are reformulated as percentile rank variables, and if log expenditure variables are employed as the dependent variables.11
Importantly, the results are also consistent in sign and magnitude if estimated unconditional on
cognitive skills and other endowment measures. There may be some concern that contemporaneous measures of cognitive skills are endogenous relative to parental investments, but there is no
evidence that including these cognitive measures as controls generates systematic bias; the results
from estimating a more parsimonious specification are reported in Panel A of Table A2.
We also re-estimate the results using two alternate samples. First, we employ the full sample of
11
These results are reported in the on-line appendix.
11
all households where the number of children is greater than or equal to two, rather than restricting
to households in which the index child and his or her younger sibling are the only children.12
(Less than 10% of the sample of interest are households with only one child.) The estimation
results can be found in Panel B of Table A2, and the interaction terms on maternal education
are again negative and and of roughly equal magnitude, suggesting that the observed pattern of
compensation in households with more educated mothers is not limited to households of a particular
size. (Households in which the index child has an older sibling and no younger siblings are still
not observed in this sample due to the absence of data on the older sibling. However, the evidence
presented in Table 1 suggests that there is no significant difference in household characteristics or
the non-cognitive or cognitive skills of the index child when comparing index children observed in
the subsample to the full sample.)13
Second, it should be noted that gender cannot be considered to be exogenous in this sample;
nearly 70% of second-born children are boys. However, consistent with existing anthropological
evidence, there is little evidence of sex selection prior to the first birth (Gu et al., 2007). The
gender ratio for first-born children is not significantly different from .5, and the gender ratio for
second-born children who follow the birth of a son is also not significantly different from .5. Thus,
sex selection is a phenomenon primarily observed after the birth of a first-born girl. Accordingly, if
we restrict the sample to households reporting a first-born son, the distribution of gender in these
households can plausibly be considered quasi-exogenous. Re-estimating the primary specification
for this smaller sample generates the same pattern of results, reported in Panel C of Table A2.
Serial correlation in parental investment
Parents may use various strategies to determine
investment between children: they may seek to maximize the returns on their educational investments, or they may allocate based on other criteria. Regardless of the allocation strategy, if there
is some serial correlation in investment, this may be a source of bias in our results.
Consider the following simple specification as an example, where the dependent variable is
educational expenditure and Yi,t−1 denotes previous parental investment.
Yit = β1 N oncogit + β2 Yi,t−1
(3)
If we assume that prior investment is unobserved and is positively correlated with expenditure
today, then Cov(N oncogit , Yi,t−1 ) > 0 will generate upward bias on the coefficient β1 . If β1 is
positive (i.e., investment is higher for children with higher non-cognitive skills), then we will over12
We preferentially employ the restricted sample in the primary results given that failing to observe the endowment
of the third sibling may lead to attenuation bias in the primary results if parents are compensating the unobserved
sibling.
13
We also explore whether there is evidence of a correlation between birth spacing (between the first- and secondborn child) and maternal education that could be an alternate channel for the detected pattern. There is no evidence
of any correlation between birth spacing and maternal or paternal education, and no evidence that parents respond
differentially to non-cognitive skills in households with different spacing between the two siblings.
12
estimate β1 and over-estimate the degree of reinforcing investment. If β1 is negative (i.e., investment
is higher for children with lower non-cognitive skills), then the coefficient will be biased toward zero,
and we will under-estimate the degree of compensatory investment. The bias in the coefficient can
be captured by the familiar bias term for omitted variables,
β2 Cov(Yi,t−1 ,N oncogit )
.
V ar(N oncogit )
β2 captures the
degree of serial correlation in investment.
The question of primary interest for our specification is whether there will be bias in the
interaction term between non-cognitive skills and maternal education – and, if the bias does exist,
why do we observe this pattern only for maternal education, but not paternal education?14 If we
estimated equation (3) separately for low-education and high-education mothers and then calculated
the difference in the estimated correlations β1 as a proxy for the interaction term, the bias would
cancel unless
β2 Cov(Yi,t−1 ,N oncogit )
V ar(N oncogit )
is different for the two samples. This could arise if either the
variance in non-cognitive skills or the degree of serial correlation in investment is different for more
and less educated mothers (i.e., β2 is not the same), or if the relationship between past investment
and current non-cognitive skills is different for more and less educated mothers (i.e., the returns to
investment are not the same).
We can present some evidence on these points using data reported on non-cognitive skills and
expenditure for the first-born child over time. In general, there is positive serial correlation in
investment that seems to be significantly larger in families with more educated mothers.15 However,
there is no evidence that the dependence of current non-cognitive skills on past investment differs
significantly in households with more or less educated mothers, and there is also no evidence that
V ar(N oncogit ) ais statistically different between the two groups. Thus if anything, the upward
bias on β1 would be larger in the sample of more educated mothers, generating upward bias on the
interaction term of interest in the main specification; this is, of course, in the opposite direction of
the observed effect.
We can also perform an additional test to evaluate the potential for bias in the primary specifications due to serial correlation in investment. We presume that there is limited scope for parental
investment (via educational expenditure or other, correlated measures) to affect non-cognitive skills
prim
prior to primary school. Accordingly, we construct a new variable N cogihct
that is defined as non-
cognitive skills at primary school age, as observed in wave one for the older sibling or wave two
prim
for the younger. We similarly define expenditure at primary school age Yihct
. We then estimate
a specification parallel to the specification of interest, though rather than using year-of-birth fixed
effects, we employ fixed effects for the age of the child in the survey year, νage .
prim
prim
prim
Yihct
= β1 N cogihct
+ β2 N cogihct
× Shct + β3 Xihct + νage + ηh + ihct
(4)
14
We will also present results in Section 4.2 that there is no evidence of comparable heterogeneity with respect to
household income.
15
More specifically, the correlation between Yi,t−1 and Yit is 0.22 for children with mothers above the median level
of education, and 0.11 for children with mothers below the median level of education.
13
The results from estimating equation (4) are reported in Panels A and B of Table A3. We
observe a similar pattern to the primary results: evidence of greater compensation in households
with more educated mothers, and no systematic variation with respect to paternal education. While
the coefficients are smaller in magnitude, the mean levels of expenditure are also lower when we
examine only children at primary school age. The estimated coefficients suggest that a one standard
deviation increase in maternal education leads to an increase in discretionary expenditure directed
to a child with one standard deviation lower non-cognitive skills of around 27%, compared to an
effect size of 35% using the original specification. This result is consistent with the hypothesis that
the observed pattern does not solely reflect bias introduced by differential prior investment.
It is also possible to re-estimate our primary specification controlling directly for lagged expenditure; this is analogous to attempting to solve the omitted variable bias challenge captured in
equation (3) by directly including the omitted variable. The primary results, reported in Panel C
Table A3, are all consistent in both magnitude and significance when re-estimated including lagged
measures of expenditure.
Parental favoritism
Parental favoritism may interact with non-cognitive skills in two ways.
First, if a child with better non-cognitive skills is more likely to be the parental favorite and thus
receives more parental investment, favoritism could be interpreted as a channel through which noncognitive skills affect educational investment. Second, if parents choose a favorite child early in
life, invest more in this favored child, and he is then observed to have better non-cognitive skills,
then parental favoritism could be the underlying reason for a pattern of serial correlation — and
if parental favoritism and the associated persistence in investment is more prevalent in certain
households, this could be a source of bias in our main specification.
We implement two tests to evaluate whether favoritism is relevant in this analysis. First, we
exploit questions in the survey of mothers in which she reports the identity of the children who will
provide more emotional and economic support in the future. We designate a child as a favorite if he
is identified as the primary source of both emotional and economic support, and examine whether
more educated mothers are more likely to provide more expenditure to their favorite children,
conditional on non-cognitive skills and the full set of child characteristic controls already reported.
Importantly, there is no significant correlation between non-cognitive skills and the identity of the
favorite child, nor is there any heterogeneity in this correlation with respect to maternal education,
suggesting that it is not the case that children with stronger non-cognitive skills are more likely to
build relationships with parents and be identified as the favorite.
Second, we evaluate whether parents invest more in a favored child, and if this pattern is
different in households with more educated mothers. The specification of interest is as follows,
where Fihct denotes a dummy for favorite.
Yihct = β1 N cogihct + β2 N cogihct × Shct + β3 Fihct + β4 Fhct × Shct + β5 Xihct + νt + ηh + ihct
(5)
14
The results are reported in Table A4. In general, there is no systematic preference in expenditure
for the favored child; there is some evidence of greater preference for the favored child among more
educated mothers, but the difference is not significant.16 These results suggest that favoritism is
unlikely to be an important channel for the observed pattern of compensation.
Alternate measures of non-cognitive skills Our primary measure of non-cognitive skills is
based on children’s self-reported status. This may raise questions about noise in the data, especially
given that some children in the sample are relatively young. In particular, it is possible that children
from some households are better able to understand and respond to the questions designed to elicit
measures of their non-cognitive skills; this could result, for example, in a pattern in which there is
greater measurement error in non-cognitive skills in households with less educated mothers.
The other measure of non-cognitive skills available in this data is drawn from surveys of the
children’s teachers, who are asked to report whether a child possesses a series of eight characteristics,
both positive and negative.17 Given the limited variation in this teacher-reported measure, which
has only eight unique values, we construct a dummy variable equal to one if the teacher’s reports
D .
of the child’s characteristics places him or her in the top half of all children, denoted T cogihct
We then estimate a specification parallel to our primary specification of interest, employing this
dummy variable as a measure of non-cognitive skills.
D
D
× Shct + β3 Xihct + νt + ηh + ihct
+ β2 T cogihct
Yihct = β1 T cogihct
(6)
The results are reported in Table A5 and show the same pattern of greater compensation in
households with more educated mothers. The magnitude observed is similar to the magnitude in
a parallel specification using a dummy variable constructed from the original self-reported noncognitive skills measure.
This evidence suggests that the observed pattern of differential compensation for adverse noncognitive skills in households with more educated mothers is not an artifact of the non-cognitive
self-reports. However, we preferentially use the measures of non-cognitive skills derived from selfreports given that these measures have greater variation, and only self-reported measures can be
employed in the longitudinal analysis discussed in Section 5. Teacher reports are, clearly, no longer
available for the index children once they have completed secondary school.
16
It is also useful to note that the favorite child is significantly more likely to be a boy, though there is no significant
variation in this relationship with respect to maternal education.
17
This series includes whether the child is smart, conscientious, reasonable/well-mannered, clean, enjoys work, is
lively/imaginative, gets along with others, likes to cry, or lacks confidence.
15
4.2
Channels
There are several channels that would be consistent with the observed pattern of compensation only
in high maternal education households. First, more educated mothers may simply be better able to
learn about variation in non-cognitive skills among their children, while less educated mothers do
not acquire this information. (By contrast, more educated fathers are no better able to recognize
differences in non-cognitive skills compared to less educated fathers.) Second, maternal education
may be a proxy for income or socioeconomic status, and higher-income households may have a
preference for compensatory investment.
Third, more educated women may have a greater preference for compensatory investment relative to less educated women; alternatively, all women may have a greater preference for compensatory investment, while only educated women are able to impose this preference within the
household. A gender difference in preferences would be consistent with the finding from the experimental literature that women are in general more inequality-averse (Andreoni and Vesterlund, 2001; Dickinson and Tiefenthaler, 2002; Selten and Ockenfels, 1998; Dufwenberg and Muren,
2006), with the caveat that we cannot rule out that compensatory investment in this context is a
returns-maximizing strategy if returns to investment are higher for children with lower non-cognitive
skills.18 In addition, a large literature has presented evidence that women exhibit a preference for
greater investment in health and education for children, (Bobonis, 2009; Duflo, 2003; Fafchamps
and Quisumbing, 2002; Quisumbing and Maluccio, 2003); women or more educated women may
also have a stronger preference for compensatory investment. We will explore each channel in turn.
Parental learning about child characteristics
The first postulated channel is that more
educated mothers are better able to recognize deficits in non-cognitive skills among their children
and respond appropriately, while less educated mothers do not recognize deficits in non-cognitive
skills. It is also possible that more educated mothers acquire information about the returns to
investing in children with weak non-cognitive skills that less educated mothers do not acquire. In
either case, it is also necessary to assume there is no correlation between paternal education and
the relevant information acquisition process.
If this channel is important, then ceteris paribus we would expect more compensation among
parents who spend more time with their children, and thus have the opportunity to learn more
about their child’s characteristics and/or how these characteristics affect the returns to investment
on that child. In this survey, parents report the time spent in a typical week either playing with or
talking to children; about 65% of both mothers and fathers report that in a typical week they spend
any time on either of these activities, and the mean time spent is around three hours, only slightly
higher for mothers than fathers. (Importantly, they do not report the division of this time between
18
For example, Andreoni and Vesterlund (2001) find that women are more concerned about equalizing earnings
during experiments, while men are more focused on maximizing efficiency.
16
children, and thus we cannot analyze the time spent with each child as a dependent variable.) Also,
similar to the evidence presented in Guryan et al. (2008), we find a positive, albeit relatively weak,
correlation between parental education and time spent with children.
In order to test whether the observed pattern in fact reflects a process in which educated
mothers spend more time with their children and thus learn more about their characteristics, we
estimate the following specification including the original interaction term between non-cognitive
skills and parental education, and adding an interaction term between non-cognitive skills and
average hours spent per week with children T imehct for both the mother and the father. (In order
to enable comparison of the coefficients on the two variables, in each specification the time variable
is re-scaled to have the same mean as the parental education variable.) The same control variables
included in the primary specification are included.
Yihct = β1 N cogihct + β2 N oncogihct × Shct + β3 N oncogihct × T imehct + β4 Xihct
(7)
+ νt + ηh + ihct
The results can be found in Panels A and B of Table 5 for maternal and paternal education
respectively. We can observe that the coefficients β2 remain significant and negative for the mother.
While the estimated coefficients β3 are negative, indicating there is some evidence that mothers
who spend more time with their children do compensate more, they are small in magnitude and
generally insignificant. However, it is also important to note that if the most important information
educated mothers acquire is information about variation in the returns to investment with respect
to child characteristics, rather than child characteristics themselves, this process may be orthogonal
to time spent with the child. In that case, the test implemented here would have limited power.
Maternal education as a proxy for income The second postulated channel that could be
consistent with the observed pattern of differential compensation is that maternal education is a
proxy for household income, and higher-income parents prefer compensatory investment whereas
lower-income households prefer reinforcing investment. As suggested by Almond and Currie (2011)
and Conley (2008), low socioeconomic status households could focus on better endowed children due
to credit constraints; alternatively, the elasticity of substitution between consumption and human
capital investment could be higher for low SES households. These hypotheses are somewhat less
plausible in our context given that there are no parallel effects for paternal education, and in the
sample of interest there is no evidence that maternal education is more closely correlated with
household income when compared to paternal education.
However, a simple test of this hypothesis can be implemented by adding an interaction term
between household income and non-cognitive skills to the primary specification. This yields the
following equations, where Ihct denotes household income; income and maternal education are
17
re-scaled to have the same mean in order to allow for comparison of the coefficient magnitudes.
Yihct = β1 N cogihct + β2 N cogihct × Shct + β3 Inihct × Ihct + β4 Xihct + νt + ηh + ihct
(8)
The results of estimating equation (8) employing the education of the mother as the measure
of parental education can be found in Table 6. It is evident that the coefficients β3 on the income
interaction terms are small in magnitude, inconsistent in sign, and generally insignificant. The
coefficients on the maternal education interaction terms, by contrast, remain consistently negative
and significant. We also re-estimate this specification employing indices of productive assets and
durable household goods owned by the household, rather than income. We observe some evidence
of greater compensation among higher-asset households; importantly, however, the coefficient β2
remains negative, significant, and of consistent magnitude. This suggests it is unlikely the primary
results reflect shifts in parental preferences in higher-income households.
Maternal education and preferences for compensatory investment
Finally, we seek to
evaluate whether there is evidence consistent with the hypothesis that maternal education is correlated with a greater preference for compensatory investment, or a greater ability to impose a
certain preference within the household (i.e., greater bargaining power). We do not have any data
that would enable us to identify variation in preferences for compensatory investment between more
and less educated mothers, or for mothers compared to fathers. However, we do have data on who
makes certain decisions within the household (as reported by mothers), and we employ this data to
construct two measures of household decision-making power. The first is a general decision-making
index ranging from zero to seven, corresponding to the number of decision-making categories in
which the mother reports involvement.19 The second is an index capturing whether the mother
makes decisions related to parenting and children’s education, ranging from zero to two.
We then regress these decision-making measures on both the mother’s education and the difference in education in two simple specifications, including household control variables (net income,
per-capita income, and age of both parents) and county fixed effects. The decision-making variables
are denoted Dechct .
m
Dechct = β1 Shct
+ κc + Xhct + hct
Dechct =
dif
β1 Shct
+ κc + Xhct + hct
(9)
(10)
The results are reported in Panel A of Table 7, and suggest that there is generally a positive
correlation between maternal education and the mother’s reported decision-making power.20 We
19
These include decisions about children’s schooling, purchase of durable goods, choice of crops, purchase of livestock, managing family finances, parenting methods, and household management.
20
Interestingly, the comparable correlation between paternal education and paternal bargaining power is insignificant.
18
then test whether greater bargaining power is correlated with greater compensation by re-estimating
the primary specification, adding an interaction term between the non-cognitive index and a dummy
variable capturing the mother’s reported decision-making power around parenting decisions; the
dummy variable is defined equal to one if she reports any decision-making power. The specification
of interest can thus be written as follows, where DecD
hct denotes the decision-making dummy.
Yihct = β1 N oncogihct + β2 N oncogihct × Shct + β3 N oncogihct × DecD
hct + β4 Xihct
(11)
+ νt + ηh + ihct
The results are reported in Panel B of Table 7. While the coefficients β2 on maternal schooling
remain significant and similar in magnitude compared to the coefficients reported in Table 4, the
interpretation is somewhat different. For example, the results in Table 4 suggest we would observe compensatory behavior in a household where the mother has five years of education. Here,
the results suggest if the mother has five years of education but no decision-making power (the
decision-making dummy equals zero), we would observe a reinforcing allocation of expenditure. In
contrast, if the mother is educated and has some decision-making power, the allocation of educational expenditure would reverse, favoring the child with weaker non-cognitive skills. This example
illustrates the importance of the mother’s bargaining power in conjunction with her education level
in determining the intrahousehold allocation of investment.
Since the coefficients on maternal education remain relatively unchanged, this also suggests
that maternal education may affect investment through a channel other than bargaining power.
We interpret this as evidence that different preferences among more educated women are also
important. A differential preference for compensation may reflect a belief among more educated
women that they have the knowledge to utilize expenditure to address non-cognitive deficits among
their children, while low-educated women may not believe they have this knowledge. This difference
is in addition to the implied difference in average preferences between mothers and fathers; the
latter may reflect a lack of interest by fathers in compensation, or perhaps a belief that some other
compensatory strategy is more appropriate.
These results also raise the question of whether it is solely the level of maternal education
that affects observed decisions about child investment, or the difference between maternal and
paternal education. If we re-estimate the primary results restricting the sample to households
where parents are relatively close in education (i.e., excluding households falling in the top and
bottom 25% of the distribution of the within-household educational gap), we observe that more
educated women still report greater decision-making power, and households with more educated
women still show evidence of greater compensation. In other words, high-education women married
to high-education men exhibit more compensatory behavior than low-education women married to
low-education men. This suggests that the level of education is important, perhaps because it is
19
correlated with different maternal preferences.21
However, we can also re-estimate our primary specification by replacing the interaction of
maternal education and non-cognitive skills with the interaction of the difference in education
(maternal minus paternal) and non-cognitive skills, and we also observe greater compensation
in households where the difference in education is more positive. In other words, low-education
women married to low-education men show more compensatory behavior than low-education women
married to high-education men, presumably because they can exert greater bargaining power within
the household. This evidence, in conjunction with the results presented in Panel A of Table 7,
suggests that the difference in education (relative bargaining power) is also relevant.22
Both the difference in education and the level of education seem important in determining
parents’ joint decisions. This is consistent with higher education for mothers leading to a greater
preference for compensation and more bargaining power among educated women.
It may also be useful at this point to briefly return to the evidence for the parental response
to cognitive skills. The results, though noisy, suggested that there was uniform reinforcement with
respect to cognitive skills, regardless of the level of maternal education.23 One hypothesis that
would be consistent with this pattern is that, as argued above, fathers generally have a preference
for reinforcing expenditure with respect to all dimensions of human capital for their children, and
mothers are primarily responsive to non-cognitive skills (and do not respond to cognitive skills).
Accordingly, we observe a pattern of reinforcing expenditure except with respect to non-cognitive
skills in households with more educated mothers, who have a stronger preference for compensatory
investment in this dimension, and/or the bargaining power to impose this preference. However,
given the imprecision in the estimated response to cognitive skills, this hypothesis must be regarded
as merely speculative.
5
Catch-up in non-cognitive skills over time
Given the evidence about compensatory investment in households with educated mothers, it is
plausible to hypothesize that over time, children of educated mothers with worse non-cognitive
skills should begin to catch up relative to their peers (assuming, of course, that there are in fact
positive returns to educational expenditure). In other words, the persistence over time of non21
These results, as well as the results referred to below, are reported in the on-line appendix.
The fact that the difference in education is also correlated with greater compensation also suggests that it is
unlikely that greater compensation in higher maternal-education households reflects a pattern of assortative mating
in which high-education women choose spouses whose preferences match their own. Even comparing across households
where the woman is relatively uneducated, and thus presumably commands less power in the marriage market, we
observe that a woman who is relatively more educated within her household is able to exert a preference for greater
compensation.
23
If we estimate the parental response to a summary human capital variable including cognitive, non-cognitive
skills and height-for-age, a proxy for overall health, the results are generally noisy, and show no uniform pattern of
compensation or reinforcement.
22
20
cognitive challenges should be weaker for children of highly educated mothers.
In this data, the younger sibling is only observed once (in the second wave of the survey employed
here), while the older sibling is observed in all three survey waves. Accordingly, to test whether
catch-up in non-cognitive skills is more evident in more educated households, we examine whether
the longitudinal correlation in child characteristics for the older child is weaker in households with
a more educated mother.24 More specifically, we regress various measures of non-cognitive skills
observed in 2009 and 2004 on earlier measures of non-cognitive skills for the same child. In 2009, the
psychometric measures include a Rosenberg index of self-esteem, and an index of depression.25 In
2004, internalizing and externalizing indices are reported as already noted; in 2000, the internalizing
and externalizing indices are reported, as well as a self-esteem measure.
In the psychology literature, it is also common to use rank-order measures for personality
traits (Shiner and Caspi, 2003; Roberts and DelVecchio, 2000). This is particularly relevant when
employing longitudinal data in which subjects are observed at very different ages. Accordingly, for
this analysis we convert all the non-cognitive measures into percentile measures ranking the child
with respect to children of the same gender and the same age group; the child with the strongest
non-cognitive skills is assigned the highest percentile of 1.26
These measures of non-cognitive skills will be denoted P sychihct for child i in household h in
county c born in year t, and the superscript will indicate the year in which the data was observed.
Thus the primary equations of interest can be written as follows, regressing non-cognitive outcomes
on outcomes from the previous wave and the interaction of the previous outcome with a householdlevel input that can lead to catch up in non-cognitive skills, Ihct . Ihct can be a dummy variable for
the mother or father having a high level of education (above the median), or reported discretionary
educational expenditure on the child in the previous wave. The specifications of interest are written
as follows.
2004
2004
2009
× Ihct
+ β2 N cogihct
= β1 N cogihct
N cogihct
(12)
+ β4 Xhct + νt + κc + ihct
2004
2000
2000
N cogihct
= β1 N cogihct
+ β2 N cogihct
× Ihct
(13)
+ β4 Xhct + νt + κc + ihct
Xihct denotes a vector of child- and household-level controls, and standard errors are clustered
24
Another test that could be implemented is to examine whether the absolute difference in human capital characteristics between the first-born and second-born children is narrower in wave two in households with more educated
mothers. This would be consistent with compensatory investment already successfully leading to “catch-up” by the
weaker sibling. This test shows no significant differences in the absolute difference comparing across households with
more or less educated mothers. Results are reported in the on-line appendix.
25
We invert the depression index, such that a higher depression index is indicative of a lower level of depression.
26
The primary results are similar when estimated employing the original variables, though more noisily estimated.
21
at the county level. The control variables of interest are drawn from the same set of covariates
employed in the earlier analysis: cognitive test scores as measured in 2000 and 2004, math and
Chinese test scores as measured in 2004, height-for-age as measured in 2004, household net income,
fixed capital and assets as measured in 2004, paternal and maternal education dummies, the number
of siblings in the family, gender, gender of the younger sibling, sibling gender interacted with the
number of siblings, and county and year-of-birth fixed effects. We also include an interaction term
with household net income as measured in 2004. In the specifications including an expenditure
interaction effect, additional controls include linear and quadratic terms for total and discretionary
educational expenditure, and a dummy for discretionary expenditure above the median.
The results of estimating equations (12) and (13) for maternal and paternal education dummies and educational expenditure are reported in Table 8. Note a positive coefficient β1 can be
interpreted as evidence of persistence of non-cognitive skills over time, and a negative coefficient
β2 can be interpreted as catch-up in households with higher levels of parental education or more
educational expenditure. The interaction terms with maternal and paternal education are included
in the same specification.
First, it is useful to note that non-cognitive skills at ages 9–12 (as measured in 2000) do not
seem to be particularly strongly correlated with non-cognitive skills at ages 13–16 (as measured in
2004); β1 is positive, but not always statistically significant. There appears to be greater evidence
of persistence between ages 13–16 and young adulthood, or between 2004 and 2009.
Second, and more importantly, there is also evidence of catch-up in non-cognitive skills for
children of more educated mothers (as reported in Columns 1, 3, 5, and 7), and for children who
receive more educational expenditure (as reported in Columns 2, 4, 6, and 8). This is observed
both between 2000 and 2004 and between 2004 and 2009, and is consistent with compensatory
investment by mothers facilitating catch-up by children with weaker non-cognitive skills.27 This
pattern is consistent with the existing literature suggesting non-cognitive skills are more malleable
for a longer period into adolescence and young adulthood than cognitive skills (Borghans et al.,
2008). (It is, however, important to be cautious in interpreting these results as evidence of returns
to the specific investments observed: more educated mothers may also make additional, unobserved
investments targeting children with weaker non-cognitive skills that leads to catch-up.)
The interaction terms with paternal education, by contrast, are generally positive and insignificant. Given that there was little evidence that more educated fathers compensated children of
lower non-cognitive skills with additional investment, this is consistent with our prior results.
An alternate test that captures the same fundamental empirical pattern examines the crosshousehold correlation between non-cognitive skills and a dummy variable for the household being
characterized by high maternal education, conditional on the same set of control variables. This
27
There is no evidence of comparable catch-up in cognitive skills. This is unsurprising, given that we did not
observe a compensatory response among households with more educated mothers in response to variation in children’s
cognitive skills.
22
correlation is increasing in magnitude in each wave: in the first wave, the observed cross-household
correlation between a dummy for a mother of high maternal education and the percentile measures
of non-cognitive skills is essentially zero. In the second wave, the correlation has increased in
magnitude to .028, and by the third wave, .151. The difference between the first- and third-wave
coefficients is statistically significant at the five percent level.28 There is no evidence of a comparable
pattern for paternal education.
Given this pattern, it is also useful to briefly reconsider our primary results of heterogeneous
response to variation in non-cognitive skills with respect to parental education. The longitudinal
evidence suggests that in households with more educated mothers, compensatory investments may
already have succeeded in generating some catch-up in non-cognitive skills among first-born children
with weaker non-cognitive skills prior to wave two. However, we have already presented evidence
in section 4.1 (Table A3) that the primary results are robust to employing the initial, wave-one
measure of non-cognitive skills for the older child. In addition, any catch-up prior to wave two would
lead to a narrowing of the gap in non-cognitive skills between children of an educated mother, and
thus lead us to underestimate the compensatory behavior engaged in by these mothers. There is
no obvious source of bias that would lead us to erroneously conclude that educated mothers are
compensating when in fact they are reinforcing.
Considering the long-term effect of the observed patterns, we can again compare a child with
a mother below the median level of education to a child with a mother above the median level of
education. For the former child, a one standard deviation decrease in the non-cognitive percentile
rank in adolescence leads to a .23 decrease in the Rosenberg percentile rank in young adulthood
(i.e., the child has lower self-esteem).29 For the latter child, however, the same decrease in the noncognitive index in adolescence does not lead to any statistically significant change in self-esteem
in adulthood. If there are positive returns in the labor market to non-cognitive skills such as
self-esteem, this pattern may have meaningful economic implications.
In addition, these results may raise the question of why parents appear to utilize educational
expenditure as a tool to address deficits in non-cognitive skills as opposed to some other form of
expenditure more explicitly targeted at developing these skills. We should note that we cannot
rule out that parents are simultaneously making other, targeted investments, in expenditure or in
time, to develop non-cognitive skills, and that the observed pattern represents primarily returns to
these unobserved investments. As previously highlighted, educational expenditure is the only type
of child-specific investment reported in this survey other than medical care. However, given that
these are relatively resource-constrained households, forms of expenditure that might be considered
appropriate for developing non-cognitive skills in a developed county (e.g., therapeutic interven28
The difference between the coefficients for the first and second wave is not significant, and the difference between
the coefficients for the second and third wave is narrowly insignificant at the ten percent level.
29
We assume the father is also above the median level of education, and thus the relevant coefficient is the sum of
.110 and .118 as reported in Column (5) of Table 8.
23
tions, additional attention from teachers, or intense supervision by parents) may be unavailable.
In fact, both parents report spending only around four hours total per week with their children
playing, talking, or assisting with homework, and this includes time spent with both children.
If we examine the evidence about categories of expenditure that respond to non-cognitive skills,
referring back to Table 3, we observe the largest effects for food at school, transportation to school,
and tutoring. Tutoring for children after school hours may be a viable strategy to ensure that they
engage in constructive activities when the parents are unavailable for supervision. Expenditure on
food at school and transportation to school may correspond to a decision to allow the child to spend
more time at school or attend a school that is farther away from home, enhancing exposure to peers
and teachers and potentially enhancing non-cognitive skills. Needless to say, we also cannot rule
out that these investments lead to simultaneous enhanced development of cognitive skills.
6
Conclusion
The decisions parents make about how to allocate educational investments among children have
major implications for policies targeting human capital accumulation. As greater emphasis is placed
on the development of non-cognitive skills as well as cognitive skills as a strategy for increasing
long-term welfare, it is even more important to understand how parents may respond to observed
differences in non-cognitive skills, and whether they seek to address any detectable deficits.
The evidence in this paper suggests that in a rural developing country context, households
with more educated mothers may engage in more compensatory behavior, targeting expenditure to
children with weak non-cognitive skills, when compared to households with less educated mothers.
While the observed pattern would be consistent with many channels and we must be cautious in
interpreting our findings, the evidence presented here suggests that both greater bargaining power
for more educated mothers and a differential preference for compensation among more educated
mothers are relevant. Over time, this leads to greater persistence in non-cognitive challenges in
households with less educated mothers, while children in households with more educated mothers
show evidence of catch-up.
One implication is that maternal education plays a significant role in child development. There
is an extensive literature that find the intergenerational transmission of the mother’s education and
one’s education achievement(Black et al., 2005b).30 Our finding suggests part of the intergenerational transmission of education may work through the impact of maternal education on children’s
non-cognitive abilities, which in terms affect the child’s education outcome.
Given that Gansu is one of the poorest regions of China—our sample is characterized by per
capita income of only around $200—our results suggest that compensatory behavior by parents
can be found even in resource-constrained settings. This pattern also may have implications for
30
See Björklund et al. (2011) for a review of this literature.
24
interventions targeted to strengthen non-cognitive skills. If more educated mothers respond to such
interventions by redirecting expenditure away from the child whose skills have been strengthened,
this may be a mechanism that decreases the long-term benefits for the targeted child. There may,
however, be positive spillovers for other siblings.
It is important to note that our sample is relatively small and drawn from only one province
in China, and we cannot conclude that the observed phenomenon is a general one. However, our
results suggest that the question of whether differential household responses to child variation in
non-cognitive skills widen cross-household inequality in human capital over time may merit further
analysis.
25
References
Adhvaryu, Achyuta and Anant Nyshadham, “Endowments at birth and parents’ investments
in children,” The Economic Journal, 2014.
, Teresa Molina, Anant Nyshadham, and Jorge Tamayo, “Helping Children Catch Up:
Early Life Shocks and the Progresa Experiment,” 2015.
Aizer, Anna and Flavio Cunha, “The production of human capital: Endowments, investments
and fertility,” Working Paper 18429, National Bureau of Economic Research September 2012.
Akresh, Richard, Emilie Bagby, Damien de Walque, and Harounan Kazianga, “Child
ability and household human capital investment in Burkina Faso,” Economic Development and
Cultural Change, 2012, 61 (1), 157–186.
Almond, Douglas and Bhashkar Mazumder, “Fetal origins and parental responses,” Annu.
Rev. Econ., 2013, 5 (1), 37–56.
and Janet Currie, “Human capital development before age five,” Handbook of Labor Economics, 2011, 4, 1315–1486.
, Hilary Hoynes, and Diane Schanzenbach, “Childhood Exposure to the Food Stamp Program: Long-run Health and Economic Outcomes,” American Economic Review, forthcoming.
, Lena Edlund, and Maarten Palme, “Chernobyl’s subclinical legacy: Prenatal exposure to
radioactive fallout and school outcomes in Sweden,” Quarterly Journal of Economics, 2009, 124
(4), 1729–1772.
Andreoni, James and Lise Vesterlund, “Which is the fair sex? Gender differences in altruism,”
Quarterly Journal of Economics, 2001, pp. 293–312.
Bhalotra, Sonia and Atheendar Venkataramani, “Shadows of the captain of the men of
death: Early life health, human capital investment and institutions.,” 2016. Mimeo.
Bharadwaj, Prashant, Juan Eberhard, and Christopher Neilson, “Health at birth, parental
investments and academic outcomes,” University of California at San Diego Working Paper
Series, 2013.
, Katrine Vellesen Løken, and Christopher Neilson, “Early life health interventions and
academic achievement,” American Economic Review, 2013, 103 (5), 1862–1891.
Björklund, Anders, Kjell G Salvanes et al., “Education and Family Background: Mechanisms
and Policies,” Handbook of the Economics of Education, 2011, 3, 201–247.
26
Black, Sandra E, Paul J Devereux, and Kjell G Salvanes, “The more the merrier? The
effect of family size and birth order on children’s education,” The Quarterly Journal of Economics,
2005, pp. 669–700.
,
, and
, “Why the apple doesn’t fall far: Understanding intergenerational transmission of
human capital,” The American economic review, 2005, 95 (1), 437–449.
,
, and
, “Small family, smart family? Family size and the IQ scores of young men,” Journal
of Human Resources, 2010, 45 (1), 33–58.
Bleakley, Hoyt, “Malaria Eradication in the Americas: A Retrospective Analysis of Childhood
Exposure,” American Economic Journal: Applied Economics, 2010, 122 (1), 73–117.
Bobonis, Gustavo, “Is the allocation of resources within the household efficient? New evidence
from a randomized experiment,” Journal of Political Economy, 2009, 117 (3), 453–503.
Bono, Emilia Del, John Ermisch, and Marco Francesconi, “Intrafamily resource allocations:
a dynamic structural model of birth weight,” Journal of Labor Economics, 2012, 30 (3), 657–706.
Borghans, Lex, Angela Duckworth, James Heckman, and Bas ter Weel, “The economics
and psychology of personality traits,” Journal of Human Resources, 2008, 43 (4), 972–1059.
Carneiro, Pedro, Claire Crawford, and Alissa Goodman, “The impact of early cognitive
and non-cognitive skills on later outcomes,” 2007.
, Costas Meghir, and Matthias Parey, “Maternal education, home environments, and the
development of children and adolescents,” Journal of the European Economic Association, 2013,
11 (s1), 123–160.
Conley, Dalton, “Bringing sibling differences in: Enlarging our understanding of transmission of
advantage in families,” Social Class: How Does It Work?, 2008, p. 179.
Cunha, Flavio and James Heckman, “The technology of skill formation,” American Economic
Review P & P, 2007, 97 (2), 31–47.
,
, and Suzanne Schennach, “Estimating the elasticity of intertemporal substitution in the
formation of cognitive and non-cognitive skills,” Econometrica, 2010, 78 (3), 883–931.
,
, Lance Lochner, and Dimitriy Masterov, “Interpreting the evidence on life cycle skill
formation,” in Eric A. Hanushek, Stephen Machin, and Ludger Woessmann, eds., Handbook of
the Economics of Education, Amsterdam: Elsevier, 2006, pp. 697–812.
Currie, Janet, “Healthy, wealthy, and wise: Is there a causal relationship between child health
and human capital development?,” Journal of Economic Literature, 2009, 47 (1), 87–122.
27
Dickinson, David L. and Jill Tiefenthaler, “What is fair? Experimental evidence,” Southern
Economic Journal, 2002, pp. 414–428.
Duflo, Esther, “Grandmothers and granddaughters: Old-age pensions and intrahousehold allocation in South Africa,” The World Bank Economic Review, 2003, 17 (1), 1–25.
Dufwenberg, Martin and Astri Muren, “Gender composition in teams,” Journal of Economic
Behavior & Organization, 2006, 61 (1), 50–54.
Fafchamps, Marcel and Agnes Quisumbing, “Control and ownership of assets within rural
Ethiopian households,” Journal of Development Studies, 2002, 38, 47–82.
Frijters, Paul, David Johnston, Manisha Shah, and Michael Shields, “Intrahousehold
resource allocation: do parents reduce or reinforce child ability gaps?,” Demography, 2013, 50
(6), 2187–2208.
Gelber, Alexander and Adam Isen, “Children’s schooling and parents’ behavior: Evidence
from the Head Start Impact Study,” Journal of Public Economics, 2013, 101, 25–38.
Glewwe, Paul, Qiuqiong Huang, and Albert Park, “Cognitive skills, non-cognitive skills and
the employment and wages of young adults in rural China,” 2013. Working paper.
Gould, Eric D., Victor Lavy, and M. Daniele Paserman, “Sixty years after the magic carpet
ride: The long-run effect of the early childhood environment on social and economic outcomes,”
Review of Economic Studies, 2011, 78, 938–973.
Gu, Baochang, Wang Feng, Guo Zhigang, and Zhang Erli, “China’s local and national
fertility policies at the end of the twentieth century,” Population and Development Review, Mar.
2007, 33 (1), 129–147.
Guryan, Jonathan, Erik Hurst, and Melissa Kearney, “Parental education and parental
time with children,” The Journal of Economic Perspectives, 2008, 22 (3), 23–46.
Heckman, James J and Yona Rubinstein, “The importance of noncognitive skills: Lessons
from the GED testing program,” American Economic Review, 2001, 91 (2), 145–149.
Heckman, James J., Jora Stixrud, and Sergio Urzua, “The effects of cognitive and noncognitive abilities on labor market outcomes and social behavior,” Journal of Labor Economics,
2006, 24, 411–482.
Hsin, Amy, “Is biology destiny? Birth weight and differential parental treatment,” Demography,
2012, 49 (4), 1385–1405.
28
Kling, Jeffrey, Jeffrey Liebman, and Lawrence Katz, “Experimental analysis of neighborhood effects,” Econometrica, 2007, 75, 83119.
Leight, Jessica, “Sibling rivalry: Ability and intrahousehold allocation in Gansu Province,
China,” 2014. Mimeo.
Quisumbing, Agnes and John Maluccio, “Resources at marriage and intrahousehold allocation: Evidence from Bangladesh, Ethiopia, Indonesia and South Africa,” Oxford Bulletin of
Economics and Statistics, 2003, 65, 283–327.
Restropo, Brandon, “Parental Investment Responses to a Low Birth Weight Outcome: Who
Compensates and Who Reinforces,” Forthcoming, Journal of Population Economics, 2016.
Roberts, Brent W. and Wendy F. DelVecchio, “The rank-order consistency of personality
traits from childhood to old age: a quantitative review of longitudinal studies,” Psychological
Bulletin, 2000, 126 (1), 3.
Rosenzweig, Mark and Junsen Zhang, “Do population control policies induce more human
capital investment? Twins, birth weight and China’s “One-Child” Policy,” Review of Economic
Studies, 2009, 76 (3), 1149–74.
Royer, Heather, “Separated at girth: US twin estimates of the effects of birth weight,” American
Economic Journal: Applied Economics, 2009, 1 (1), 49–85.
Schweinhart, Lawrence J., Jeanne Montie, Zongping Xiang, William S. Barnett,
Clive R. Belfield, and Milagros Nores, Lifetime effects: The HighScope Perry preschool
study through age 40, Ypsilanti, Michigan: Monographs of the HighScope Educational Research
Foundation, 2005.
Selten, Reinhard and Axel Ockenfels, “An experimental solidarity game,” Journal of Economic Behavior and Organization, 1998, 34 (4), 517–539.
Shiner, Rebecca and Avshalom Caspi, “Personality differences in childhood and adolescence:
Measurement, development, and consequences,” Journal of Child Psychology and Psychiatry,
2003, 44 (1), 2–32.
29
7
Figures and Tables
Figure 1: Data structure
Figure 2: Parental education
(a) Maternal education
(b) Paternal education
30
Figure 3: Intra-household differences in non-cognitive skills and expenditure, by mother’s education
(a) Non-cognitive skills
(b) Expenditure
Notes: For each level of mother’s education in years, the bottom bar corresponds to the minimum value of
between-sibling absolute differences in non-cognitive skills (Figure 3a) or between-sibling absolute differences in
normalized educational expenditure residuals (Figure 3b), while the top bar corresponds to the maximum value.
The rectangle corresponds to the interquartile range, with the median value represented by the bold line bisecting
the rectangle. Normalized expenditure residuals are calculated by regressing expenditure on child characteristics
(gender, birth parity, cognitive skills, and height-for-age), generating the residuals and standardizing them to have
mean zero and standard deviation one.
Table 1: Summary statistics
Panel A: Demographic data
Net income
Income per capita
Mother education
Father education
Mother age
Father age
Index child age
Internalizing index
Externalizing index
Achievement index
Obs.
Panel B: Educational expenditure per child
Sample
Subsample
p-value
6848.56
1717.76
4.31
7.12
39.19
42.57
15.09
-.01
.03
0
6728.28
1631.04
4.35
7.16
36.97
38.89
14.96
.01
-.01
.06
.863
.581
.805
.837
.000
.072
.013
.634
.280
.183
1918
388
Total
Discretionary
Tuition
Supplies
Transportation
Food
Tutoring
Other fees
Mean
Std. Dev.
Max.
361.52
149.03
212.49
47.07
19.67
59.12
9.66
13.50
343.1
215.65
185.47
43.55
50.99
139.08
21.42
36.41
2900
1660
2000
300
500
1200
110
360
Notes: The sample encompasses the full sample of households that report income data; this is 1914 out of the full
sample of 2000 households in the survey. The subsample is households with two-children families in which both
children report data on non-cognitive skills as well as height-for-age. There are 388 households in the subsample of
interest, and 776 children. Income is reported in yuan; educational expenditure is reported in yuan per semester.
Internalizing, externalizing and achievement indices have been standardized to have means equal to zero and
standard deviations equal to one. A higher internalizing or externalizing index is indicative of higher non-cognitive
skills. Column (3) reports the p-value for a test of equality of means across the sample and subsample.
31
32
776
776
.042
(.039)
.013
(.059)
(.030)
-.118∗∗
.003
(.058)
External
(2)
(.018)
.315
(.083)
-.006
(.077)
∗∗∗
External
(6)
(.003)
776
.002
(.005)
-.0007
776
776
.005
(.005)
.0009
(.005)
Panel B: Cognitive skills and health
.039∗∗
.008
776
Internal
(5)
Panel A: Child characteristics
External
(4)
(.017)
Internal
(3)
776
776
-.005
(.003)
(.003)
(.020)
(.019)
-.006∗∗
.068∗∗∗
External
(8)
.015
Internal
(7)
776
(.003)
-.006∗∗
(.007)
.004
(.005)
-.0005
(.029)
.019
(.045)
.046
(.080)
-.007
(.051)
.007
(.140)
.143
Internal
(9)
776
(.003)
-.005∗
(.007)
.007
(.006)
-.0009
(.038)
.045
(.054)
.159∗∗∗
(.074)
.289∗∗∗
(.044)
-.076∗
(.119)
.131
External
(10)
Notes: The dependent variables are the internalizing and externalizing indices. A higher internalizing or externalizing index is indicative of higher
non-cognitive skills. The independent variable is the specified child characteristic, all measured in the second wave; all specifications include household
fixed effects and standard errors clustered at the county level. Asterisks indicate significant at the 10, 5 and 1 percent levels.
Obs.
Achievement score
Chinese score
Math score
Height-for-age
Grade level
Female
Age
Sibling parity
Internal
(1)
Table 2: Non-cognitive skills and child characteristics
Table 3: Parental allocations and non-cognitive skills
Total
(1)
Discretionary
(2)
Tuition
(3)
Supplies
(4)
Transportation
(5)
Food
(6)
Tutoring
(7)
Other
(8)
Panel A: Internalizing index
Index
-8.010
-10.019
2.009
.591
-.617
-10.713∗
-1.056
1.776
(10.770)
(8.021)
(4.065)
(1.635)
(1.247)
(5.873)
(.712)
(1.437)
Panel B: Externalizing index
Index
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
.028
-2.772
2.800
2.366∗
.252
-6.048
-.284
.941
(11.850)
(7.854)
(5.576)
(1.405)
(1.348)
(5.799)
(.730)
(.912)
776
776
776
776
776
776
776
776
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variables are educational expenditure per semester per child in the specified category; total
expenditure is the sum of all categories, and discretionary expenditure is the sum of all categories excluding tuition.
The independent variable is the specified index of internalizing or externalizing behavior. A higher internalizing or
externalizing index is indicative of higher non-cognitive skills. All specifications include controls for sibling parity,
gender, sibling gender, height-for-age and cognitive skills measured contemporaneously with non-cognitive skills,
year-of-birth and household fixed effects, and standard errors clustered at the county level. Asterisks indicate
significance at the 10, 5 and 1 percent levels.
33
Table 4: Heterogeneous effects with respect to parental education
Total
(1)
Discretionary Tuition Supplies Transportation
(2)
(3)
(4)
(5)
Food
(6)
Tutoring
(7)
Other
(8)
Panel A: Internalizing index and maternal education
Index
27.973
20.386
7.588
.392
4.365∗
10.157
1.253
4.219∗∗
(17.338)
(14.144)
(7.585)
(1.600)
(2.504)
(11.232)
(.829)
(2.080)
∗∗
∗∗
Index x mother educ. -8.595
(4.186)
-7.262
(3.617)
-1.333
.048
(1.686)
(.426)
∗∗
-1.190
(.606)
∗
-4.985
(2.652)
∗∗
-.551
-.583
(.231)
(.442)
Panel B: Internalizing index and paternal education
Index
Index x father educ.
Test β2m = β2f
Joint test
-11.252
-17.317
6.066
-2.838
5.045∗
-9.825
-2.712
-6.987∗∗∗
(20.718)
(16.862)
(8.148)
(1.905)
(2.664)
(14.485)
(1.725)
(2.676)
.286
.949
-.663
.488
-.131
.204
1.163∗∗
(2.329)
(1.580)
(1.414)
(.314)
(.340)
(1.355)
(.170)
(.472)
.072
.184
.023
.073
.847
.857
.430
.484
.589
.084
.070
.180
.017
.052
.013
.034
-.775
∗∗
Panel C: Externalizing index and maternal education
Index
Index x mother educ
36.951∗∗
22.555∗
14.396∗
-.033
5.606∗∗
13.883
1.764∗∗
1.335
(15.425)
(11.563)
(8.706)
(1.281)
(2.391)
(8.492)
(.813)
(1.227)
-9.971∗∗
-6.840∗
-3.131
.648
-1.446∗∗
-5.382∗∗
-.553∗
-.106
(4.740)
(3.531)
(2.495)
(.403)
(.644)
(2.531)
(.309)
(.325)
Panel D: Externalizing index and paternal education
Index
Index x father educ
Test β2m = β2f
Joint test
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
-18.195
-23.448
5.253
-.946
1.977
-18.834
-2.200
-3.444∗
(26.820)
(19.075)
(11.498)
(2.617)
(2.666)
(15.555)
(1.857)
(2.008)
2.644
2.855
-.211
.408
-.197
1.758
.243
.643∗
(2.725)
(1.983)
(1.287)
(.325)
(.362)
(1.570)
(.203)
(.347)
.027
.085
.020
.058
.344
.453
.601
.168
.064
.099
.028
.089
.051
.144
.347
.473
766
276.748
286.515
766
104.052
174.218
766
172.696
161.381
766
38.164
38.785
766
12.011
40.065
766
37.679
109.107
766
6.218
17.357
766
9.98
28.532
Notes: The dependent variables are educational expenditure per semester per child in the specified category; total
expenditure is the sum of all categories, and discretionary expenditure is the sum of all categories excluding tuition.
The independent variable is the specified index of internalizing or externalizing behavior, as well as the index
interacted with the specified measure of parental education in years. A higher internalizing or externalizing index is
indicative of higher non-cognitive skills. All specifications include controls for sibling parity, gender, sibling gender,
height-for-age and cognitive skills measured contemporaneously with non-cognitive skills, sibling parity, gender,
height-for-age and achievement scores interacted with maternal education, year-of-birth and household fixed effects,
and standard errors clustered at the county level. Panels B and D report tests of equality of the interaction term β2
across parallel specifications employing maternal and paternal education, as well as joint tests of the equalities
β1m = β1f and β2m = β2f . Asterisks indicate significance at the 10, 5 and 1 percent levels.
34
Table 5: Non-cognitive skills and parental learning
Total
(1)
Discretionary
(2)
Tuition
(3)
Supplies
(4)
Transp.
(5)
Food
(6)
Tutoring
(7)
Other
(8)
Panel A: Non-cognitive index and maternal time investment
Index
Index x mother educ.
Index x time
56.650∗∗∗
39.536∗∗∗
17.114∗
2.472
6.391∗∗
23.109∗∗
2.845∗∗
4.718∗∗
(17.674)
(14.165)
(9.529)
(1.943)
(2.554)
(10.783)
(1.365)
(2.170)
-11.198∗∗
-8.597∗∗
-2.601
.447
-1.645∗∗
-6.325∗∗
-.677∗∗
-.396
(5.085)
(4.292)
(2.088)
(.453)
(.702)
(3.138)
(.301)
(.433)
-3.765
-2.925
-.840
-.524
-.035
-1.836
-.218
-.312∗∗
(2.716)
(2.401)
(.658)
(.353)
(.289)
(2.112)
(.159)
(.155)
Panel B: Non-cognitive index and paternal time investment
Index
Index x father educ.
Index x time
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
-21.437
-26.774
5.337
-1.285
2.899
-20.299
-3.369
-4.721∗∗
(32.034)
(23.348)
(12.583)
(3.316)
(2.563)
(19.510)
(2.354)
(2.054)
1.678
2.183
-.504
.511
-.575
.914
.269
1.065∗∗
(2.772)
(1.856)
(1.496)
(.356)
(.356)
(1.612)
(.214)
(.431)
.487
.333
.154
-.097
.146
.398
.048
-.161
(.828)
(.511)
(.466)
(.231)
(.118)
(.405)
(.086)
(.108)
766
766
766
766
766
766
766
766
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variables in Panels A and B are educational expenditure per semester per child in the
specified category; total expenditure is the sum of all categories, and discretionary expenditure is the sum of all
categories excluding tuition. The independent variables are the specified index of internalizing or externalizing
behavior, the index interacted with the specified measure of parental education in years, and the index interacted
with the amount of time the parent reports spending with children playing, talking or assisting with homework in a
typical week. A higher non-cognitive index is indicative of higher non-cognitive skills. Time invested by the mother
(father) is re-scaled to have the same mean as education of the mother (father). All specifications include controls
for sibling parity, gender, sibling gender, height-for-age and cognitive skills measured contemporaneously with
non-cognitive skills, sibling parity, gender, age, height-for-age and achievement scores interacted with maternal
education, year-of-birth and household fixed effects, and standard errors clustered at the county level. Standard
errors are clustered at the level of the county. Asterisks indicate significance at the 10, 5 and 1 percent levels.
35
Table 6: Heterogeneous effects with respect to maternal education and income
Total
(1)
Index
Index x educ.
Index x income
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
Discretionary
(2)
Tuition
(3)
Supplies
(4)
Transportation
(5)
Food
(6)
Tutoring
(7)
Other
(8)
42.020∗∗
29.815∗
12.204
.520
6.639∗∗
16.680
2.264∗∗
3.712∗∗
(19.792)
(15.645)
(8.968)
(1.473)
(3.053)
(11.959)
(.891)
(1.841)
-11.308∗∗
-8.161∗∗
-3.146
.458
-1.524∗∗
-6.156∗∗
-.599∗
-.340
(4.879)
(4.124)
(2.185)
(.522)
(.616)
(3.030)
(.307)
(.465)
-.571
-1.243
.672
-.120
-.192
-.621
-.162∗
-.148
(1.401)
(.896)
(.678)
(.149)
(.198)
(.563)
(.087)
(.111)
776
776
776
776
776
776
776
776
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variables are educational expenditure per semester per child in the specified category; total
expenditure is the sum of all categories, and discretionary expenditure is the sum of all categories excluding tuition.
The independent variable is the specified index of internalizing or externalizing behavior, the index interacted with
the specified measure of parental education in years, and the index interacted with household net income. A higher
non-cognitive index is indicative of higher non-cognitive skills. Net income is re-scaled to have the same mean as
maternal education. All specifications include controls for sibling parity, gender, sibling gender, height-for-age and
cognitive skills measured contemporaneously with non-cognitive skills, sibling parity, gender, age, height-for-age and
achievement scores interacted with maternal education, year-of-birth and household fixed effects, and standard
errors clustered at the county level. Asterisks indicate significance at the 10, 5 and 1 percent levels.
36
Table 7: Maternal education and bargaining power
Panel A: Reported household decision-making by mother
(1)
Dif educ.
All decisions
(2)
.036∗
.004
(.019)
(.006)
Mother educ.
Mean
Obs.
Parenting decisions
(3)
(4)
2.952
383
.051∗∗
.017∗∗
(.025)
(.008)
2.952
388
.857
383
.857
388
Panel B: Non-cognitive index and maternal decision-making
Index
Total
Discretionary Tuition
76.364∗∗∗
62.670∗∗∗
13.694
(23.839)
Index x mother educ.
-11.429
∗∗
(5.175)
Index x decision-making -40.159∗
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
(19.364)
∗∗
-8.721
(4.316)
(14.853)
Supplies
.845
(2.838)
Transportation
Food
Tutoring Other
14.829∗∗
38.894∗∗ 6.160∗∗∗ 1.942
(5.984)
∗∗
-2.708
.384
(2.086)
(.465)
-1.589
(.703)
(15.355)
(2.038)
∗∗
∗∗
-6.394
(3.129)
-.675
(.313)
(2.032)
-.446
(.452)
-40.040∗∗∗
-.119
-.647
-9.707∗
(23.719)
(15.173)
(15.661)
(2.857)
(5.891)
-26.561∗∗ -4.782∗∗∗ 1.657
(13.077)
(1.708)
(2.047)
776
776
776
776
776
776
776
776
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variable in Panel A, Columns (1) and (3) is an index indicating the number of decisions
made by the mother, as reported by the mother (ranges from 0 to 7); the dependent variable in Columns (4) and
(6) is an index indicating the number of decisions made by the father, as reported by the father (ranges from 0 to
7). The independent variables in Panel A are the difference between maternal education and paternal education in
years, maternal education in years or paternal education in years, as specified, and controls included are household
net income, per-capita income, maternal and paternal age, and county fixed effects.
In Panel B, the dependent variables are educational expenditure as defined in Table 3, and the independent
variables are the non-cognitive index, the index interacted with a dummy variable equal to one if the mother
reports any decision-making power around parenting, and the index interacted with maternal education. A higher
non-cognitive index is indicative of higher non-cognitive skills. Specifications in Panel B include controls for sibling
parity, gender, sibling gender, height-for-age and cognitive skills measured contemporaneously with non-cognitive
skills, sibling parity, gender, age, height-for-age and achievement scores interacted with maternal education,
year-of-birth and household fixed effects, and standard errors clustered at the county level. Asterisks indicate
significance at the 10, 5 and 1 percent levels.
37
Table 8: Persistence of non-cognitive skills in percentile and parental education
Internal 2004
(1)
(2)
Psychometric index 2000
Mother educ. 2000 int.
Father educ. 2000 int.
External 2004
(3)
(4)
.001
.067
-.005
.046
(.093)
(.078)
(.059)
(.054)
-.143∗∗
-.083
(.059)
(.070)
.159
.127
(.121)
(.106)
Exp. 2000 int.
-.002∗∗∗
-.0009∗∗
(.0005)
(.0004)
Psychometric index 2004
Mother educ. 2004 int.
Father educ. 2004 int.
Rosenberg 2009
(5)
(6)
.110
.093
.131∗∗
.164∗∗∗
(.096)
(.070)
(.063)
(.048)
∗
-.181
-.004
(.107)
(.099)
.118
-.025
(.116)
(.106)
Exp. 2004 int.
Obs.
550
550
550
550
Depression 2009
(7)
(8)
408
-.00002
-.0004∗∗
(.0003)
(.0002)
408
410
410
Notes: The dependent variables are the specified measure of non-cognitive skills from waves two and three,
calculated in percentile terms. A higher index in internalizing or externalizing behavior is indicative of higher
non-cognitive skills. A higher percentile in the Rosenberg index is associated with higher self-esteem. The
depression index is inverted, and thus a higher percentile in the depression index is indicative of a lower level of
depression. The independent variables include the psychometric index, the mean of internalizing and externalizing
indices from waves one and two, calculated in percentile terms. The psychometric index is also interacted with
dummy variables for the mother’s (father’s) education being above the median, and discretionary educational
expenditure. All specifications include controls for cognitive skills measured in waves one and two, grades in math
and Chinese, height-for-age, the parental education dummy variables, net income, assets, and fixed capital in the
household, the number of siblings, sibling gender, and sibling gender interacted with the number of siblings, and
county and year-of-birth fixed effects; the specifications including an interaction effect with expenditure also include
linear and quadratic terms for total and discretionary educational expenditure, and a dummy for discretionary
expenditure above the median. Standard errors are clustered at the level of the county. Asterisks indicate
significance at the 10, 5 and 1 percent levels.
38
Appendix
Table A1: Within-household and cross-household variation
Internal
(1)
(2)
Fixed effects
Obs.
R-squared
External
(3)
(4)
Height-for-age
(5)
(6)
Achievement score
(7)
(8)
Grades
(9)
(10)
Household
Year
Household
Year
Household
Year
Household
Year
Household
Year
776
.600
776
.021
776
.620
776
.016
776
.828
776
.067
776
.864
776
.054
776
.676
776
.034
Notes: Each column reports the R-squared for the regression of the specified child characteristic on the specified set
of fixed effects.
39
Table A2: Parental allocations and non-cognitive skills: Alternate specifications
Total
(1)
Discretionary
(2)
Tuition
(3)
Supplies Transportation
(4)
(5)
Food
(6)
Tutoring
(7)
Other
(8)
Panel A: Results unconditional on child characteristics
Index
38.182∗∗
24.918∗
13.265
.520
5.902∗∗
13.422
1.465∗∗
3.609
(15.827)
(13.257)
(8.286)
(1.443)
(2.552)
(10.643)
(.720)
(2.584)
Index x mother educ. -11.467∗∗
Obs.
-8.740∗∗
-2.727
.129
-1.534∗∗∗
-6.293∗∗
-.525∗
-.516
(4.752)
(3.939)
(2.169)
(.528)
(.552)
(2.829)
(.284)
(.519)
776
776
776
776
776
776
776
776
Panel B: Households including two or more children
Index
Index x mother educ.
Obs.
29.733∗∗
18.451∗
11.282∗∗
.786
4.462∗
9.526
1.042
2.635∗∗
(12.185)
(9.588)
(5.406)
(.965)
(2.334)
(7.144)
(.965)
(1.119)
-8.238∗∗
-5.680∗∗
-2.558
.269
-.864
-3.756∗
-.999
-.330
(3.554)
(2.896)
(1.576)
(.335)
(.540)
(1.973)
(.668)
(.275)
1132
1132
1132
1132
1132
1132
1132
1132
Panel C: Households including a first-born son
Index
Index x mother educ.
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
48.862∗
39.267∗∗
9.595
1.790
6.373
27.235∗∗
2.954∗∗
.915
(25.905)
(18.256)
(11.187)
(1.876)
(3.928)
(13.106)
(1.444)
(1.593)
-13.411∗
-11.645∗∗
-1.766
-.184
-1.674
-8.777∗∗
-1.022∗∗
.012
(7.697)
(5.729)
(2.911)
(.520)
(1.077)
(4.148)
(.437)
(.669)
390
390
390
390
390
390
390
390
261.867
273.416
94.244
168.236
167.622
148.407
36.22
37.226
10.658
37.467
31.111
102.985
7.025
33.987
9.23
30.367
Notes: The dependent variables are educational expenditure per semester per child in the specified category; total
expenditure is the sum of all categories, and discretionary expenditure is the sum of all categories excluding tuition.
The independent variable is a mean index of non-cognitive skills and the interaction of this index with maternal
education. A higher internalizing or externalizing index is indicative of higher non-cognitive skills. In Panel A, the
specifications include year-of-birth and household fixed effects, and standard errors clustered at the county level; the
sample is households with two children. In Panel B, all specifications include controls for sibling parity, gender,
sibling gender, height-for-age and cognitive skills, sibling parity, gender, age, height-for-age and achievement scores
interacted with maternal education, year-of-birth and household fixed effects, and standard errors clustered at the
county level; the sample is households with two or more children. Asterisks indicate significance at the 10, 5 and 1
percent levels.
40
Table A3: Heterogeneous effects: Addressing bias due to serial correlation in expenditure
Total
(1)
Discretionary Tuition Supplies Transp. and food Tutoring Other fees
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Panel A: Maternal education and non-cognitive index
Index
Index x mother educ.
Obs.
9.480∗
6.518∗
2.962
1.455
1.611
.317
.069
(5.508)
(3.862)
(3.558)
(1.904)
(1.090)
(.286)
(1.448)
-3.820∗∗
-2.643∗∗
-1.177
-.897∗∗
-1.045∗
-.115
-.054
(1.829)
(1.208)
(1.061)
(.420)
(.582)
(.082)
(.464)
756
756
756
756
756
756
756
Panel B: Paternal education and non-cognitive index
Index
Index x father educ.
Mean (dep. var.)
St. dev. (dep. var.)
Obs.
-17.016
3.775
-20.791
1.076
-2.584
-.781∗∗
2.551
(18.593)
(7.605)
(15.264)
(2.487)
(2.235)
(.379)
(2.340)
1.762
-1.017
2.779
-.423
.048
.096∗
-.445
(2.450)
(.770)
(2.122)
(.311)
(.236)
(.056)
(.346)
159.95
133.65
756
49.19
78.21
756
110.76
89.91
756
25.06
24.35
756
5.42
32.7
756
1.52
7.9
756
9.75
25.11
756
Panel C: Maternal education including controls for past expenditure
Total
Index
Transportation
Food
Tutoring
Other
40.791∗∗
27.190∗
13.601
.267
6.230∗∗
15.364
1.927∗∗
3.402∗
(18.974)
(15.689)
(8.923)
(1.399)
(2.987)
(12.149)
(.955)
(1.822)
Index x mother educ. -11.692∗∗
Mean (dep. var.)
St. dev. (dep. var.)
Obs.
Discretionary Tuition Supplies
-8.987∗∗
-2.705
.378
-1.653∗∗
-6.569∗∗
-.706∗∗
-.437
(5.004)
(4.238)
(2.105)
(.463)
(.707)
(3.095)
(.312)
(.433)
276.748
286.515
776
104.052
174.218
776
172.696
161.381
776
38.164
38.785
776
12.011
40.065
776
37.679
109.107
776
6.218
17.357
776
9.98
28.532
776
Notes: In Panel A and Panel B, the dependent variables are educational expenditure per semester per child in the
specified category at primary school age (as observed in wave one for the first-born child and in wave two for the
second-born child); total expenditure is the sum of all categories, and discretionary expenditure is the sum of all
categories excluding tuition. The independent variables in Panels A and B include a non-cognitive index as
measured at primary school age, as well as the index interacted with the specified measure of parental education. A
higher internalizing or externalizing index is indicative of higher non-cognitive skills. All specifications include
controls for sibling parity, gender, sibling gender, height-for-age and cognitive skills measured contemporaneously
with non-cognitive skills, sibling parity, gender, age, height-for-age and achievement scores interacted with maternal
education, year-of-birth and household fixed effects, and standard errors clustered at the county level. Specifications
in Panel C are similar to those in Table 4 and include an additional control for lagged educational expenditure (as
reported in wave one). Asterisks indicate significance at the 10, 5 and 1 percent levels.
41
Table A4: Heterogeneous effects with respect to maternal favorite
Total
Index
Index x mother educ.
Favorite
Favorite x mother educ.
Obs.
Mean (dep. var.)
St. dev. (dep. var.)
Discretionary Tuition Supplies Transportation
Food
Tutoring Other
40.644∗∗
27.199∗
13.445
.283
6.230∗∗
15.348
1.902∗∗
3.435∗
(18.599)
(15.198)
(8.873)
(1.370)
(2.845)
(11.894)
(.926)
(1.846)
-11.693∗∗
-9.023∗∗
-2.670
.366
-1.653∗∗
-6.587∗∗
-.701∗∗
-.447
(5.181)
(4.352)
(2.111)
(.460)
(.702)
(3.163)
(.320)
(.448)
11.545
-1.244
12.789
-1.940
.535
.839
2.113
-2.790
(24.908)
(19.064)
(9.275)
(6.472)
(4.874)
(13.812)
(2.238)
(1.912)
6.370
7.054
-.684
1.222
.723
4.281
.293
.535
(8.962)
(5.643)
(4.555)
(1.476)
(1.375)
(4.232)
(.490)
(.520)
776
776
776
776
776
776
776
776
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variable is educational expenditure in the specified category, and the independent variables
are the non-cognitive index and a dummy for the child being reported as the maternal favorite, and both variables
interacted with maternal education. A higher internalizing or externalizing index is indicative of higher
non-cognitive skills. All specifications include controls for sibling parity, gender, sibling gender, height-for-age and
cognitive skills measured contemporaneously with non-cognitive skills, sibling parity, gender, age, height-for-age and
achievement scores interacted with maternal education, year-of-birth and household fixed effects, and standard
errors clustered at the county level. Standard errors are clustered at the level of the county.
Table A5: Alternate measures of non-cognitive skills: teacher reports
Total
(1)
Teacher index
Mean (dep. var.)
St. dev. (dep. var.)
Food
(6)
Tutoring
(7)
Other
(8)
34.132
28.784∗
5.348
8.086
4.973
13.544
.136
2.045
(25.439)
(16.150)
(12.351)
(5.030)
(4.655)
(8.644)
(2.061)
(2.044)
Index x mother educ. -14.294∗∗
Obs.
Discretionary Tuition Supplies Transportation
(2)
(3)
(4)
(5)
-13.210∗∗∗
-1.085
-1.586∗∗
-2.785∗∗
-8.496∗∗∗
-.282
-.061
(5.642)
(4.078)
(2.626)
(.764)
(1.121)
(2.551)
(.592)
(.272)
774
774
774
774
774
774
774
774
276.748
286.515
104.052
174.218
172.696
161.381
38.164
38.785
12.011
40.065
37.679
109.107
6.218
17.357
9.98
28.532
Notes: The dependent variables are educational expenditure per semester per child in the specified category; total
expenditure is the sum of all categories, and discretionary expenditure is the sum of all categories excluding tuition.
The independent variables are a dummy variable equal to one if an index of non-cognitive skills based on teacher
reports of the child’s behavior is above the mean, and the interaction of this dummy variable with maternal
education. All specifications include controls for sibling parity, gender, sibling gender, height-for-age and cognitive
skills measured contemporaneously with non-cognitive skills, sibling parity, gender, age, height-for-age and
achievement scores interacted with maternal education, year-of-birth and household fixed effects, and standard
errors clustered at the county level. Asterisks indicate significance at the 10, 5 and 1 percent levels.
42